Precision-recall Curve Research Articles

DEEPFAKE DETECTION USING TRANSFER LEARNING-BASED XCEPTION MODEL

Justification of the purpose of the research. In recent times, several approaches for face manipulation in videos have been extensively applied and availed to the public which makes editing faces in video easy for everyone effortlessly with realistic efforts. While beneficial in various domains, these methods could significantly harm society if employed to spread misinformation. So, it is also vital to properly detect whether a face has been distorted in a video series. To detect this deepfake, convolutional neural networks can be used in past works. However, it needs a greater number of parameters and more computations. So, to overcome these limitations and to accurately detect deepfakes in videos, a transfer learning-based model named the Improved Xception model is suggested. Obtained results. This model is trained using extracted facial landmark features with robust training. Moreover, the improved Xception model's detection accuracy is evaluated alongside ResNet and Inception, considering model loss, accuracy, ROC, training time, and the Precision-Recall curve. The outcomes confirm the success of the proposed model, which employs transfer learning techniques to identify fraudulent videos. Furthermore, the method demonstrates a noteworthy 5% increase in efficiency compared to current systems.

Simple Histogram Equalization Technique Improves Performance of VGG Models on Facial Emotion Recognition Datasets

Facial emotion recognition (FER) is crucial across psychology, neuroscience, computer vision, and machine learning due to the diversified and subjective nature of emotions, varying considerably across individuals, cultures, and contexts. This study explored FER through convolutional neural networks (CNNs) and Histogram Equalization techniques. It investigated the impact of histogram equalization, data augmentation, and various model optimization strategies on FER accuracy across different datasets like KDEF, CK+, and FER2013. Using pre-trained VGG architectures, such as VGG19 and VGG16, this study also examined the effectiveness of fine-tuning hyperparameters and implementing different learning rate schedulers. The evaluation encompassed diverse metrics including accuracy, Area Under the Receiver Operating Characteristic Curve (AUC-ROC), Area Under the Precision–Recall Curve (AUC-PRC), and Weighted F1 score. Notably, the fine-tuned VGG architecture demonstrated a state-of-the-art performance compared to conventional transfer learning models and achieved 100%, 95.92%, and 69.65% on the CK+, KDEF, and FER2013 datasets, respectively.

RNA m6A detection using raw current signals and basecalling errors from Nanopore direct RNA sequencing reads.

Nanopore direct RNA sequencing (DRS) enables the detection of RNA N6-methyladenosine (m6A) without extra laboratory techniques. A number of supervised or comparative approaches have been developed to identify m6A from Nanopore DRS reads. However, existing methods typically utilize either statistical features of the current signals or basecalling-error features, ignoring the richer information of the raw signals of DRS reads. Here, we propose RedNano, a deep-learning method designed to detect m6A from Nanopore DRS reads by utilizing both raw signals and basecalling errors. RedNano processes the raw-signal feature and basecalling-error feature through residual networks. We validated the effectiveness of RedNano using synthesized, Arabidopsis, and human DRS data. The results demonstrate that RedNano surpasses existing methods by achieving higher area under the ROC curve (AUC) and area under the precision-recall curve (AUPRs) in all three datasets. Furthermore, RedNano performs better in cross-species validation, demonstrating its robustness. Additionally, when detecting m6A from an independent dataset of Populus trichocarpa, RedNano achieves the highest AUC and AUPR, which are 3.8%-9.9% and 5.5%-13.8% higher than other methods, respectively. The source code of RedNano is freely available at https://github.com/Derryxu/RedNano.

Bayesian Network Inference for Low-Magnitude Nonnatural Seismic Event Discrimination

Abstract In response to the gaps in understanding the causal relationship between seismic waveform features and the types of seismic events, this research is focused on seismic events of low magnitude (ML≤3.0) in the North China region. Using the Bayesian network theory, we conduct an analysis to infer event types for natural earthquakes, artificial explosions, and mining collapses, and the outcomes achieved notable efficacy for the discrimination of seismic events. Through the analysis of seismic waveforms from 1818 events, we systematically extracted and quantified 55 features in temporal, spectral, and energy domains, which were then recoded as node variables for subsequent analysis. The new data set was subject to select nodes with strong associations to the node type. Subsequently, Bayesian network topologies were constructed using three different algorithms to reconstruct the custom network, calculating posterior probabilities and marginal probabilities. Simultaneously, an extensive evaluation with precision–recall curves of the network structure was carried out, encompassing accuracy, precision, recall, and F1-score. Ultimately, sensitivity analysis was performed on each node to reveal the extent of the influence of node variations on the inference of the node type. The findings showed that the sensitivity of discrimination of seismic events was notably high for several features, including high-frequency P/S spectral ratio values (11 to ∼20 Hz), central frequency, dominant frequency, average frequency, rise and decay average frequency, the real part of the complex cepstral coefficients, peak ground acceleration, and zero crossing. In the classification of natural earthquakes, artificial explosions, and mining collapses, it was observed that the probability of mining collapses was maximized when peak ground acceleration was less than 1526.08, and concurrently, the P/S spectral ratio (11 to ∼20 Hz) fell within the range of −0.25 to −0.02.

Machine learning-based model for predicting outcomes in cerebral hemorrhage patients with leukemia

Background and purposeIntracranial hemorrhage (ICH) in leukemia patients progresses rapidly with high mortality. Limited data are available on imaging studies in this population. The study aims to develop prediction models for 7-day and short-term mortality risk based on the non-contrast computed tomography (NCCT) image features. MethodsThe NCCT image features of ICH in 135 leukemia patients between 2007–2023 were retrospectively extracted using manual assessment and radiomics methods. After multiple imputation of missing laboratory data, univariate logistic regression and least absolute shrinkage and selection operator (LASSO) were used for feature selection. Random forest models were built with comprehensive evaluation and ranking of feature importance. Result135 and 129 patients were included in the studies for 7-day and short-term prognostic models, respectively. The median age of all enrolled patients was 35 years, and there were 86 male patients (63.7 %). Clinical models (validation: AUC [area under the curve] = 0.78, AUPRC [area under the precision-recall curve] = 0.73; AUC = 0.84, AUPRC = 0.86), radiomics models (validation: AUC = 0.82, AUPRC = 0.78; AUC = 0.75, AUPRC = 0.77), and the combined models (validation: AUC = 0.84, AUPRC = 0.83; AUC = 0.87, AUPRC = 0.89) predicted 7-day and short-term mortality with good predictive efficacy. Clinical decision curve analysis showed that the combined models predicted 7-day and 30-day risk of death would be more beneficial than other models. Shape features contributed significantly more than semantic features in both radiomics models and combined models (93.3 %, 52.1 %, as well as 85.2 %,37.4 %, respectively) for 7-day and 30-day mortality. ConclusionsCombined models constructed based on NCCT perform well in predicting the risk of 7-day and short-term mortality in ICH patients with leukemia. Shape features extracted by radiomics are important markers for modeling the prognosis.

Personalized glycemic management for patients with diabetic ketoacidosis based on machine learning

To explore the optimal blood glucose-lowering strategies for patients with diabetic ketoacidosis (DKA) to enhance personalized treatment effects using machine learning techniques based on the United States Critical Care Medical Information Mart for Intensive Care- IV (MIMIC- IV). Utilizing the MIMIC- IV database, the case data of 2 096 patients with DKA admitted to the intensive care unit (ICU) at Beth Israel Deaconess Medical Center from 2008 to 2019 were analyzed. Machine learning models were developed, and receiver operator characteristic curve (ROC curve) and precision-recall curve (PR curve) were plotted to evaluate the model's effectiveness in predicting four common adverse outcomes: hypoglycemia, hypokalemia, reductions in Glasgow coma scale (GCS), and extended hospital stays. The risk of adverse outcomes was analyzed in relation to the rate of blood glucose decrease. Univariate and multivariate Logistic regression analyses were conducted to examine the relationship between relevant factors and the risk of hypokalemia. Personalized risk interpretation methods and predictive technologies were applied to individualize the analysis of optimal glucose control ranges for patients. The machine learning models demonstrated excellent performance in predicting adverse outcomes in patients with DKA, with areas under the ROC curve (AUROC) and 95% confidence interval (95%CI) for predicting hypoglycemia, hypokalemia, GCS score reduction, and extended hospital stays being 0.826 (0.803-0.849), 0.850 (0.828-0.870), 0.925 (0.903-0.946), and 0.901 (0.883-0.920), respectively. Analysis of the relationship between the rate of blood glucose reduction and the risk of four adverse outcomes showed that a maximum glucose reduction rate > 6.26 mmol×L-1×h-1 significantly increased the risk of hypoglycemia (P < 0.001); a rate > 2.72 mmol×L-1×h-1 significantly elevated the risk of hypokalemia (P < 0.001); a rate > 5.53 mmol×L-1×h-1 significantly reduced the risk of GCS score reduction (P < 0.001); and a rate > 8.03 mmol×L-1×h-1 significantly shortened the length of hospital stay (P < 0.001). Multivariate Logistic regression analysis indicated significant correlations between maximum bicarbonate levels, blood urea nitrogen levels, and total insulin doses with the risk of hypokalemia (all P < 0.01). In terms of establishing personalized optimal treatment thresholds, assuming optimal glucose reduction thresholds for hypoglycemia, hypokalemia, GCS score reduction, and extended hospital stay were x1, x2, x3, x4, respectively, the recommended glucose reduction rates to minimize the risks of hypokalemia and hypoglycemia should be ≤min{x1, x2}, while those to reduce GCS score decline and extended hospital stay should be ≥ max{x3, x4}. When these ranges overlap, i.e., max{x3, x4} ≤ min{x1, x2}, this interval was the recommended optimal glucose reduction range. If there was no overlap between these ranges, i.e., max{x3, x4} > min{x1, x2}, the treatment strategy should be dynamically adjusted considering individual differences in the risk of various adverse outcomes. The machine learning models shows good performance in predicting adverse outcomes in patients with DKA, assisting in personalized blood glucose management and holding important clinical application prospects.

Predicting persistent fever in patients with cancer using machine learning.

e13552 Background: Sepsis is the leading cause of hospital readmission and death in the developed world. Cancer patients are at high risk due to their underlying disease and therapies. Meanwhile, antimicrobial resistance increases steadily, warranting a more careful administration of antibiotics. Predicting fever persistence 48 hours after initiation of antibiotics could help reduce antibiotic escalation and chest CT scans in patients where fever is predicted to subside within the next 24 hours. Methods: All cancer patients of the University Hospital Essen (UHE) between 1/1/2020-7/1/2023, receiving broad-spectrum i.v. antibiotics, and having persistent fever after 48 hours of first administration were extracted from UHE’s Fast Health Interoperability Resources clinical database. Vital signs, clinically important covariates, as well as lag features of vital signs and laboratory values were included. The binary target to predict was prevalence of body temperature ≥38 °C after 48-72 hours. Gradient boosting machines, a deep learning neural network, logistic regression (LR), random forest (RF) and a voting ensemble classifier were implemented. Hyperparameter tuning and model validation was performed through five-fold stratified cross validation (80/20% train/test split). Results: Analyzing 2.9 million patient records of UHE we obtained approx. 27,000 hospitalized adults with cancer. Restriction to hematology/oncology wards and intravenous broad-spectrum antibiotics resulted in 1,267 cancer patients. Of these, 264 patients had persistent fever after 48 hours, of which 111 patients had solid cancers, 66 lymphomas, and 87 other hematologic malignancies. 72 hours after antibiotic therapy 160 patients (60.2%) had persistent fever. Based on the area under the precision-recall curve (AUPRC) of 81.7%, the voting classifier model had the highest predictive performance. The model shows a sensitivity of 85%, a specificity of 63%, a Negative Predictive Value (NPV) of 75% and a Positive Predictive Value (PPV) of 73% on a cross-validated test set (Table). Conclusions: Machine learning-based algorithms can predict fever evolution and assist in clinical decision-making, potentially contributing to better patient care by preventing unnecessary antibiotics escalations and CT scans. Based on current model performance, 25% of chest CT scans could be obsolete as patients will not have persistent fever after 48 hours. Additional results will be presented at the meeting. [Table: see text]

Improving lung cancer health equity by applying deep learning to low dose CT screening of minority and disadvantaged patients.

1584 Background: In the US, disparities in lung cancer mortality exist for African American, Hispanic, and other minorities. Standard of care low dose CT screening (LDCT) detects early-stage disease and improves mortality, yet disparities are perpetuated in screening by eligibility criteria derived from cohorts underrepresenting these minorities. One such cohort is the National Lung Screening Trial (NLST) cohort which is 92% White. Consequently, guidelines for lung cancer screening may be insufficient to address the unique needs of diverse populations. We hypothesize that Artificial Intelligence prediction of future lung cancer risk from an individual’s LDCT can partially mitigate racial and ethnic disparities and improve health system practice guidelines by individualizing screening risk as compared to current general guidelines. Here, we benchmark a Resnet18 3D neural network trained on NLST LDCT images, Sybil, on the diverse patient population of the University of Illinois Health system (UIH) which is 20% White and 60% African American. Methods: A real-world cohort from UIH consisting of 1,450 CT studies was evaluated alongside 60,378 CT studies from the NLST cohort. All CT studies evaluated by the model were not used in model training. Using Youden’s J index as a probability cutoff, sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were evaluated. Receiver operating characteristic (ROC) and precision-recall (PR) curves were generated to assess model performance between cohorts. NLST data were truncated to achieve equivalent incidence of lung cancer with UIH when generated PR curves. Results: Multi-year prediction performance (ROC-AUC and PR-AUC) between cohorts are summarized (Table). For prediction of lung cancer within 1-year of LDCT in the UIH cohort, the model respectively demonstrated sensitivity, specificity, positive predictive value, and negative predictive value among White (0.80, 0.77, 0.54, 0.92), African American (0.84, 0.78, 0.39, 0.97) races and Hispanic (0.75, 0.73, 0.60, 0.84) and non-Hispanic (0.87, 0.77, 0.42, 0.97) ethnicities. Conclusions: Model performance was similar between the NLST (92% White) and a diverse, real-world cohort at UIH (20% White) though decreases in ROC-AUC performance in year 1 predictions and may be due to insufficient representation of minority populations during model training. Prospective studies involving larger and more representative patient populations should be conducted to further optimize the model and evaluate its clinical utility to improve lung cancer health equity in minority populations.[Table: see text]

Deep learning-based recognition of stained tongue coating images

ObjectiveTo build a dataset encompassing a large number of stained tongue coating images and process it using deep learning to automatically recognize stained tongue coating images. MethodsA total of 1 001 images of stained tongue coating from healthy students at Hunan University of Chinese Medicine and 1 007 images of pathological (non-stained) tongue coating from hospitalized patients at The First Hospital of Hunan University of Chinese Medicine with lung cancer, diabetes, and hypertension were collected. The tongue images were randomized into the training, validation, and testing datasets in a 7 : 2 : 1 ratio. A deep learning model was constructed using the ResNet50 for recognizing stained tongue coating in the training and validation datasets. The training period was 90 epochs. The model’s performance was evaluated by its accuracy, loss curve, recall, F1 score, confusion matrix, receiver operating characteristic (ROC) curve, and precision-recall (PR) curve in the tasks of predicting stained tongue coating images in the testing dataset. The accuracy of the deep learning model was compared with that of attending physicians of traditional Chinese medicine (TCM). ResultsThe training results showed that after 90 epochs, the model presented an excellent classification performance. The loss curve and accuracy were stable, showing no signs of overfitting. The model achieved an accuracy, recall, and F1 score of 92%, 91%, and 92%, respectively. The confusion matrix revealed an accuracy of 92% for the model and 69% for TCM practitioners. The areas under the ROC and PR curves were 0.97 and 0.95, respectively. Conclusion: The deep learning model constructed using ResNet50 can effectively recognize stained coating images with greater accuracy than visual inspection of TCM practitioners. This model has the potential to assist doctors in identifying false tongue coating and preventing misdiagnosis.

CT-based radiomics predicts CD38 expression and indirectly reflects clinical prognosis in epithelial ovarian cancer

BackgroundCluster of differentiation 38 (CD38) has been found to be highly expressed in various solid tumours, and its expression level may be associated with patient prognosis and survival. This study aimed to evaluate the prognostic value of CD38 expression for patients with epithelial ovarian cancer (EOC) and construct two computed tomography (CT)-based radiomics models for predicting CD38 expression. MethodsA total of 333 cases of EOC were enrolled from The Cancer Genome Atlas (TCGA) database for CD38-related bioinformatics and survival analysis. A total of 56 intersection cases from TCGA and The Cancer Imaging Archive (TCIA) databases were selected for radiomics feature extraction and model construction. Logistic regression (LR) and support vector machine (SVM) models were constructed and internally validated using 5-fold cross-validation to assess the performance of the models for CD38 expression levels. ResultsHigh CD38 expression was an independent protective factor (HR = 0.540) for overall survival (OS) in EOC patients. Five radiomics features based on CT images were selected to build models for the prediction of CD38 expression. In the training and internal validation sets, for the receiver operating characteristic (ROC) curve, the LR model reached an area under the curve (AUC) of 0.739 and 0.732, while the SVM model achieved AUC values of 0.741 and 0.700, respectively. For the precision-recall (PR) curve, the LR and SVM models demonstrated an AUC of 0.760 and 0.721. The calibration curves and decision curve analysis (DCA) provided evidence supporting the fitness and net benefit of the models. ConclusionsHigh levels of CD38 expression can improve OS in EOC patients. CT-based radiomics models can be a new predictive tool for CD38 expression, offering possibilities for individualised survival assessment for patients with EOC.

Homocysteine concentration in coronary artery disease and severity of coronary lesions.

Our previous study reckons that the impact of the rs1801133 variant of 5,10-methylenetetrahydrofolate reductase (MTHFR) on coronary artery disease (CAD) is possibly mediated by cardiometabolic disorder. This study is performed to verify this hypothesis. Four hundred and thirty CAD patients and 216 CAD-free individuals were enrolled in this case-control study. The rs1801133 variant was genotyped by PCR-RFLP. Severity of coronary lesions was evaluated by number of stenotic coronary vessels and extent of coronary stenosis. The rs1801133 T allele significantly increased homocysteine levels in patients with CAD and CAD-free individuals. Individuals with the T allele of rs1801133 had an increased risk of developing CAD. In contrast, individuals with the TT genotype of rs1801133 were at high risk of multiple vessel lesions. The carriers of CT genotype had higher levels of systolic blood pressure (SBP), low-density lipoprotein cholesterol (LDL-C), and high-sensitivity C-reactive protein (hs-CRP), and lower levels of apolipoprotein A1 (APOA1) than those with CC genotype in male patients with CAD. The receiver operating characteristic (ROC) curve and precision-recall (PR) curve indicated that hyperhomocysteinemia was sensitive to predict the severity of CAD. Multivariate logistic regression revealed that homocysteine, rs1801133, age, smoking, weight, body mass index (BMI), lipoprotein(a) [Lp(a)], and hs-CRP were independent risk factors for CAD. The increased risk of CAD and severity of coronary lesions associated with rs1801133 in the Chinese Han population were attributed, at least partly, to high homocysteine levels. Hyperhomocysteinemia had a high predictive value for severe CAD or multiple vessel lesions.

Developing a Machine Learning Model to Predict 180-day Readmission for Elderly Patients with Angina.

Readmission of elderly angina patients has become a serious problem, with a dearth of available prediction tools for readmission assessment. The objective of this study was to develop a machine learning (ML) model that can predict 180-day all-cause readmission for elderly angina patients. The clinical data for elderly angina patients was retrospectively collected. Five ML algorithms were used to develop prediction models. Area under the receiver operating characteristic curve (AUROC), area under the precision recall curve (AUPRC), and the Brier score were applied to assess predictive performance. Analysis by Shapley additive explanations (SHAP) was performed to evaluate the contribution of each variable. A total of 1502 elderly angina patients (45.74% female) were enrolled in the study. The extreme gradient boosting (XGB) model showed good predictive performance for 180-day readmission (AUROC = 0.89; AUPRC = 0.91; Brier score = 0.21). SHAP analysis revealed that the number of medications, hematocrit, and chronic obstructive pulmonary disease were important variables associated with 180-day readmission. An ML model can accurately identify elderly angina patients with a high risk of 180-day readmission. The model used to identify individual risk factors can also serve to remind clinicians of appropriate interventions that may help to prevent the readmission of patients.

Predicting drug-target interactions using matrix factorization with self-paced learning and dual similarity information.

Drug repositioning (DR) refers to a method used to find new targets for existing drugs. This method can effectively reduce the development cost of drugs, save time on drug development, and reduce the risks of drug design. The traditional experimental methods related to DR are time-consuming, expensive, and have a high failure rate. Several computational methods have been developed with the increase in data volume and computing power. In the last decade, matrix factorization (MF) methods have been widely used in DR issues. However, these methods still have some challenges. (1) The model easily falls into a bad local optimal solution due to the high noise and high missing rate in the data. (2) Single similarity information makes the learning power of the model insufficient in terms of identifying the potential associations accurately. We proposed self-paced learning with dual similarity information and MF (SPLDMF), which introduced the self-paced learning method and more information related to drugs and targets into the model to improve prediction performance. Combining self-paced learning first can effectively alleviate the model prone to fall into a bad local optimal solution because of the high noise and high data missing rate. Then, we incorporated more data into the model to improve the model's capacity for learning. Our model achieved the best results on each dataset tested. For example, the area under the receiver operating characteristic curve and the precision-recall curve of SPLDMF was 0.982 and 0.815, respectively, outperforming the state-of-the-art methods. The experimental results on five benchmark datasets and two extended datasets demonstrated the effectiveness of our approach in predicting drug-target interactions.

An efficient computational framework for gastrointestinal disorder prediction using attention-based transfer learning

Diagnosing gastrointestinal (GI) disorders, which affect parts of the digestive system such as the stomach and intestines, can be difficult even for experienced gastroenterologists due to the variety of ways these conditions present. Early diagnosis is critical for successful treatment, but the review process is time-consuming and labor-intensive. Computer-aided diagnostic (CAD) methods provide a solution by automating diagnosis, saving time, reducing workload, and lowering the likelihood of missing critical signs. In recent years, machine learning and deep learning approaches have been used to develop many CAD systems to address this issue. However, existing systems need to be improved for better safety and reliability on larger datasets before they can be used in medical diagnostics. In our study, we developed an effective CAD system for classifying eight types of GI images by combining transfer learning with an attention mechanism. Our experimental results show that ConvNeXt is an effective pre-trained network for feature extraction, and ConvNeXt+Attention (our proposed method) is a robust CAD system that outperforms other cutting-edge approaches. Our proposed method had an area under the receiver operating characteristic curve of 0.9997 and an area under the precision-recall curve of 0.9973, indicating excellent performance. The conclusion regarding the effectiveness of the system was also supported by the values of other evaluation metrics.

Deep learning nomogram for predicting neoadjuvant chemotherapy response in locally advanced gastric cancer patients

PurposeDeveloped and validated a deep learning radiomics nomogram using multi-phase contrast-enhanced computed tomography (CECT) images to predict neoadjuvant chemotherapy (NAC) response in locally advanced gastric cancer (LAGC) patients.MethodsThis multi-center study retrospectively included 322 patients diagnosed with gastric cancer from January 2013 to June 2023 at two hospitals. Handcrafted radiomics technique and the EfficientNet V2 neural network were applied to arterial, portal venous, and delayed phase CT images to extract two-dimensional handcrafted and deep learning features. A nomogram model was built by integrating the handcrafted signature, the deep learning signature, with clinical features. Discriminative ability was assessed using the receiver operating characteristics (ROC) curve and the precision-recall (P-R) curve. Model fitting was evaluated using calibration curves, and clinical utility was assessed through decision curve analysis (DCA).ResultsThe nomogram exhibited excellent performance. The area under the ROC curve (AUC) was 0.848 [95% confidence interval (CI), 0.793–0.893)], 0.802 (95% CI 0.688–0.889), and 0.751 (95% CI 0.652–0.833) for the training, internal validation, and external validation sets, respectively. The AUCs of the P-R curves were 0.838 (95% CI 0.756–0.895), 0.541 (95% CI 0.329–0.740), and 0.556 (95% CI 0.376–0.722) for the corresponding sets. The nomogram outperformed the clinical model and handcrafted signature across all sets (all P < 0.05). The nomogram model demonstrated good calibration and provided greater net benefit within the relevant threshold range compared to other models.ConclusionThis study created a deep learning nomogram using CECT images and clinical data to predict NAC response in LAGC patients undergoing surgical resection, offering personalized treatment insights.Graphical abstract

Predicting atrial fibrillation recurrence after catheter ablation using an artificial intelligence-enabled electrocardiogram algorithm

Abstract Introduction Predicting atrial fibrillation (AF) recurrence post catheter ablation may help to assess procedural eligibility and determine AF management. While several predictors of AF recurrence post ablation have been established and numerous clinical risk scores have been proposed, their performance remained underwhelming and clinical utility was limited. Hence, new predictors are needed. Artificial intelligence (AI) algorithms using deep neural networks (DNN) to analyze biometrical data might generate such predictors. DNN algorithms have been developed to identify patients with AF based on a 12-lead electrocardiogram (ECG) in sinus rhythm. Whether these algorithms can function as a predictor of AF recurrence after AF ablation remains unknown. Purpose To evaluate the prediction of AF recurrence after ablation using an AI-enabled ECG algorithm trained to predict AF on an ECG in sinus rhythm. Methods This study retrospectively analyzed observational data from the DIGITOTAL study, that monitored AF recurrence with a PPG-based smartphone application in 96 subjects after AF ablation. Patients with a 12-lead ECG in SR available within a timeframe of 3 months before the ablation procedure were included in the analysis. Although all patients had a history of AF, an AF-risk score was calculated by the DNN described elsewhere.1 Results An ECG in sinus rhythm was available in 53 patients (14 women [26.4%]; mean [SD] age, 62.0 [9.7] years) out of the 96 patients followed-up in the DIGITOTAL study. Testing the DNN on the last ECG before the ablation procedure resulted in an area under the receiver operating curve (AUC) of 0.65 (95% CI, 0.49 - 0.80), and an area under the precision recall curve (AUPRC) of 0.56 (95% CI, 0.34 - 0.78). The optimal cutoff score resulted in a sensitivity of 60.0% (95% CI, 36.1% - 80.9%), specificity of 66.7% 66.7% (95% CI, 48.2% - 82.0%), accuracy of 0.64 (95% CI, 0.50 - 0.77), F1-score of 55.8% (95% CI, 38.5% - 74.4%), positive predictive value 52.2% (95% CI, 30.6% - 73.2%) and negative predictive value 73.3% (95% CI, 54.1% - 87.7%). Patients classified in the high-risk group versus low-risk group were more likely to exhibit AF recurrence up to one year after AF ablation (hazard ratio, 2.6; 95% CI, 1.1 - 6.5; P-value = 0.037). Conclusions The AI-enabled ECG algorithm, trained to predict AF on a sinus rhythm ECG, was able to predict AF recurrence after ablation with an accuracy comparable to the existing clinical risk scores. Further studies are needed to determine whether the DNN score can be used as an independent predictor and improve existing risk scores.

Utility of the Morgan Fingerprint in Structure-Based Virtual Ligand Screening.

In modern drug discovery, virtual ligand screening (VLS) is frequently applied to identify possible hits before experimental testing and refinement due to its cost-effective nature for large compound libraries. For decades, efforts have been devoted to developing VLS methods with high accuracy. These include the state-of-the-art FINDSITE suite of approaches FINDSITEcomb2.0, FRAGSITE, and FRAGSITE2 and the meta version FRAGSITEcomb that were developed in our lab. These methods combine ligand homology modeling (LHM), traditional ligand similarity methods, and more recently machine learning approaches to rank ligands and have proven to be superior to most recent deep learning and large language model-based approaches. Here, we describe further improvements to our previous best methods by combining the Morgan fingerprint (MF) with the originally used PubChem fingerprint and FP2 fingerprint. We then benchmarked FINDSITEcomb2.0M, FRAGSITEM, FRAGSITE2M, and the composite meta-approach FRAGSITEcombM. On the 102 target DUD-E set, the 1% enrichment factor (EF1%) and area under the precision-recall curve (AUPR) of FRAGSITEcomb increased from 42.0/0.59 to 47.6/0.72. This 0.72 AUPR is significantly better than that of the state-of-the-art deep learning-based method DenseFS's AUPR of 0.443. An independent test on the 81 targets DEKOIS2.0 set shows that EF1%/AUPR increases from 18.3/0.520 to 23.1/0.683. An ablation investigation shows that the MF contributes to most of the improvement of all four approaches. Thus, the MF is a useful addition to structure-based VLS.

Can prediction of acute myocardial infarction using a deep learning-enabled electrocardiogram be used in real-world practice?

Abstract Background Diagnosing acute myocardial infarction (MI) in the emergency department poses challenges. The utilization of a 12-lead electrocardiogram (ECG) is pivotal for detecting ST-segment elevation myocardial infarction (STEMI), yet enhancing ECG interpretation through a systematic algorithm could significantly improve diagnostic accuracy. Nonetheless, delayed diagnosis or misdiagnosis of STEMI is not uncommon in daily practice. Purpose This study aims to develop a fully convolutional network (FCN) based on ECG data as a diagnostic support tool and to evaluate its effectiveness in STEMI detection. Methods A retrospective cohort study involved 3,298 patients with angiographically confirmed STEMI and 37,393 control ECGs, comprising normal readings, false positives, and ECGs mimicking STEMI. The FCN underwent internal training, external validation, and testing across various data splits (80:10:10, 70:15:15, and 60:20:20). Performance evaluation on validation data included assessing the area under the receiver operating characteristic curve (AUC). Subsequently, the trained model predicted unseen data, and sensitivity, specificity, false positives, and false negatives were calculated to evaluate model performance. Results The FCN exhibited exceptional performance in STEMI detection during both internal training and external validation across different data splits. The receiver operating characteristic curves yielded AUC values of 0.995 (80:10:10), 0.994 (70:15:15), and 0.995 (60:20:20). Consistently high values were observed in precision-recall curves, with AUC values of 0.974 (80:10:10), 0.962 (70:15:15), and 0.973 (60:20:20). When utilizing this FCN model in the clinical setting to reanalyze 20 cases of previously misdiagnosed electrocardiograms, it correctly diagnosed 17 cases. Conclusion Integrating the FCN into clinical practice can notably improve the timely and accurate diagnosis of acute MI, reducing instances of delayed or erroneous STEMI diagnoses and facilitating prompt reperfusion therapy.

KBPT: knowledge-based prompt tuning for zero-shot relation triplet extraction

Knowledge representation is increasingly recognized as an effective method for information extraction. Nevertheless, numerous studies have disregarded its potential applications in the zero-shot setting. In this article, a novel framework, called knowledge-based prompt tuning for zero-shot relation triplet extraction (KBPT), was developed, founded on external ontology knowledge. This framework serves as a catalyst for exploring relation triplet extraction (RTE) methods within low-resource scenarios, warranting further scrutiny. Zero-shot setting RTE aims to extract multiple triplets that consist of head entities, tail entities, and relation labels from an input sentence, where the extracted relation labels are those that do not exist in the training set. To address the data scarcity problem in zero-shot RTE, a technique was introduced to synthesize training samples by prompting language models to generate structured texts. Specifically, this involves integrating language model prompts with structured text methodologies to create a structured prompt template. This template draws upon relation labels and ontology knowledge to generate synthetic training examples. The incorporation of external ontological knowledge enriches the semantic representation within the prompt template, enhancing its effectiveness. Further, a multiple triplets decoding (MTD) algorithm was developed to overcome the challenge of extracting multiple relation triplets from a sentence. To bridge the gap between knowledge and text, a collective training method was established to jointly optimize embedding representations. The proposed model is model-agnostic and can be applied to various PLMs. Exhaustive experiments on four public datasets with zero-shot settings were conducted to demonstrate the effectiveness of the proposed method. Compared to the baseline models, KBPT demonstrated enhancements of up to 14.65% and 24.19% in F1 score on the Wiki-ZSL and TACRED-Revisit datasets, respectively. Moreover, the proposed model achieved better performance compared with the current state-of-the-art (SOTA) model in terms of F1 score, precision-recall (P–R) curves and AUC. The code is available at https://Github.com/Phevos75/KBPT.

Predicting post-operative atrial fibrillation after cardiac surgery using an artificial intelligence-enabled electrocardiogram algorithm

Abstract Introduction Postoperative atrial fibrillation (POAF) is common after cardiac surgery and is associated with adverse outcomes. Systematic monitoring of POAF is cumbersome, specifically beyond discharge. Therefore, risk stratification may aid to identify patients at high risk of POAF and guide monitoring strategies alongside preventive measures. However, the performance of bedside risk stratification models reliant on clinical risk factors remained underwhelming, necessitating the exploration of more sophisticated models that maintain clinical applicability. Hence, artificial intelligence algorithms (AI) have been suggested to reinforce or replace clinical risk scores. As such, a deep neural network (DNN) algorithm was developed to identify patients with AF based on a 12-lead electrocardiogram (ECG) in sinus rhythm. Whether this algorithm can identify patients at high risk of POAF remains unknown. Purpose To evaluate the usability of an AI-enabled ECG algorithm, that was trained to predict AF in non-surgical conditions, for the prediction of POAF. Methods This study retrospectively analyzed data from the SURGICAL-AF trial that monitored patients after cardiac surgery. The inclusion criteria for this subanalysis comprised: (1) patients without a history of AF prior to cardiac surgery; (2) availability of the raw data of a pre-operative 12-lead ECG in sinus rhythm; and (3) patients with POAF (during hospitalization or up to 91 days after discharge) or patients having completed PPG-based rhythm monitoring per protocol. The AF-risk score was calculated by the DNN described elsewere.1 Results In total, 127 patients (mean [SD] age, 63.4 [8.4] years; 30 women [23.6%];, median [interquartile range] CHA2DS2-VASc score, 2 [1-3]) complied with the inclusion criteria, out of the 450 patients randomized in the SURGICAL-AF trial. Testing the DNN on the last ECG before cardiac surgery resulted in an area under the receiver operating curve (AUC) of 0.66 (95% CI, 0.56 - 0.77) and an area under the precision-recall curve of 0.57 (95% CI, 0.42 -0.72). The optimal cut of score resulted in a sensitivity of 64.3% (95% CI, 48.0%-78.4%), specificity of 64.7% (95% CI, 53.6%-74.8%), accuracy of 0.65 (95% CI, 0.56 - 0.73), F1-score of 54.5% (95% CI, 42.9% - 66.8%), positive predictive value of 47.4% (95% CI, 34.0%-61.0%), and negative predictive value of 78.6% (95% CI, 67.1%-87.5%). POAF occurred within three months after cardiac surgery in 23 patients out of 57 patients classified in the high-risk group (40.4%) versus 15 patients (21.4%) out of 70 patients classified in the low-risk group (hazard ratio, 2.2; 95% CI, 1.2 – 4.3; P-value = 0.020). Conclusions The AI-enabled ECG algorithm, trained to predict AF on a pre-operative sinus rhythm ECG, was able to identify POAF with an accuracy comparable to existing clinical risk scores. Further studies are needed to determine whether the DNN score can be used as an independent predictor and improve existing risk scores.